How Agent-Reach Skill Installation Works for OpenClaw, Claude Code, and .agents

Agent-Reach installs its capabilities by detecting your AI agent platform and copying a localized SKILL.md file to the appropriate directory, supporting OpenClaw, Claude Code, and generic .agents frameworks in a single command.

Agent-Reach is an open-source tool that bridges AI agents with external capabilities through a standardized skill format. When you run the installation command, the CLI automatically detects whether you are running OpenClaw, Claude Code, or a generic .agents compatible framework, then deploys the correct localization of its routing instructions to the platform-specific skill directory.

Platform Detection and Locale Resolution

The installation process begins by determining two critical variables: which AI agent platforms are present on the system, and which language localization to deploy.

Target Directory Priority

Inside agent_reach/cli.py (lines 13-17), the CLI builds an ordered list of potential skill directories. The priority order is:

1. $OPENCLAW_HOME/.openclaw/skills (if the OPENCLAW_HOME environment variable is set)
2. ~/.openclaw/skills (OpenClaw default)
3. ~/.claude/skills (Claude Code)
4. ~/.agents/skills (generic .agents framework)

The installer iterates through this list and installs to the first directory that already exists on the filesystem. If none exist, it falls back to creating ~/.agents/skills/agent-reach.

Locale-Aware Skill Selection

Before copying any files, the _skill_resource_name() function (lines 53-62) checks environment variables AGENT_REACH_LANG, LC_ALL, LC_MESSAGES, and LANG. If any value starts with "en" or "english", the CLI selects SKILL_en.md from the package resources; otherwise, it defaults to SKILL.md (Chinese).

The Installation Workflow

Once the target platform and locale are identified, the _copy_skill_dir() function (lines 71-108) handles the actual file operations.

Reading Package Resources

The _read_skill_markdown() function (lines 64-70) loads the selected markdown file from within the agent_reach/skill/ package directory using Python's importlib resources, ensuring the data is accessible regardless of how the package is installed.

Handling Existing Installations

The installer performs safe cleanup before writing new files:

• If an existing agent-reach entry is a symlink, it is unlinked
• If it is a regular directory, it is removed via shutil.rmtree
• The directory is then recreated fresh

This prevents conflicts and ensures a clean installation state.

Copying Reference Documentation

After writing the main SKILL.md file, the function iterates over the references/ package subdirectory (lines 99-105) and copies all helper markdown files into the new skill folder. These reference files contain platform-specific routing tables and usage examples that the AI agent consumes at runtime.

Platform-Specific Behavior

Each supported AI agent framework expects skills in a specific location, and Agent-Reach adapts its deployment accordingly.

OpenClaw Installation

For OpenClaw, the skill is installed to ~/.openclaw/skills/agent-reach (or $OPENCLAW_HOME/.openclaw/skills/agent-reach if the environment variable is set). OpenClaw loads any SKILL.md found under ~/.openclaw/skills/*/ and uses the routing table within to dispatch user intents to Agent-Reach's backend.

Claude Code Installation

Claude Code expects skills in ~/.claude/skills/. By installing the SKILL.md and references there, Claude Code can discover and invoke agent-reach as a backend tool for internet research tasks.

Generic .agents Fallback

The lightweight .agents framework reads skills from ~/.agents/skills/. Installing here makes the capability available to any agent following the standard "skill" contract (implementing the run_skill hook), serving as the default fallback when no other platform directories are detected.

CLI Usage Examples

Installing the Skill

To install into the first available skill directory:

agent-reach skill --install

Typical output on a system with OpenClaw:

Skill installed for OpenClaw: /home/user/.openclaw/skills/agent-reach

Resulting directory structure:

~/.openclaw/
└─ skills/
   └─ agent-reach/
      ├─ SKILL.md
      └─ references/
          ├─ search.md
          └─ social.md

Forcing English Localization

Override the default locale detection to install the English skill version:

export AGENT_REACH_LANG=en_US
agent-reach skill --install

The CLI copies SKILL_en.md as SKILL.md in the target directory.

Uninstalling the Skill

Remove the skill from all detected platform directories:

agent-reach skill --uninstall

The _uninstall_skill() function (lines 44-78) checks all known locations—OpenClaw, Claude Code, and .agents—and removes the agent-reach folder or symlink from each.

Summary

• Multi-platform detection: The CLI checks for OpenClaw, Claude Code, and .agents directories in priority order, defaulting to .agents if none exist.
• Locale awareness: Environment variables AGENT_REACH_LANG, LC_ALL, LC_MESSAGES, and LANG determine whether SKILL_en.md or the default SKILL.md is deployed.
• Safe installation: The _copy_skill_dir() function handles existing symlinks and directories by unlinking or recursively removing them before writing fresh files.
• Reference bundling: The references/ subdirectory is copied alongside the main skill file to provide AI agents with complete routing documentation.
• Pure Python: All operations use standard library file manipulation (shutil, pathlib) without invoking external processes.

Frequently Asked Questions

What file does Agent-Reach actually install?

Agent-Reach installs a SKILL.md markdown file containing routing instructions and triggers, plus a references/ directory with helper documentation. The specific content comes from either SKILL.md (default Chinese) or SKILL_en.md (English) stored in the package's agent_reach/skill/ directory.

Can I install the skill for multiple agents at once?

Yes. The installation loop in agent_reach/cli.py (lines 24-32) iterates through all existing platform directories and installs the skill to each one. If you have both OpenClaw and Claude Code directories present, running agent-reach skill --install will deploy to both locations simultaneously.

How do I switch between English and Chinese skill versions?

Set the AGENT_REACH_LANG environment variable to an English locale code (e.g., en, en_US) before running the install command, or ensure LC_ALL or LANG starts with "en". The _skill_resource_name() function detects this and selects SKILL_en.md instead of the default file.

What happens if I run the install command multiple times?

The installer is idempotent. The _copy_skill_dir() function removes any existing agent-reach directory or symlink before creating a fresh copy, ensuring the latest skill version is deployed without leaving orphaned files or conflicting symlinks behind.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}